solar-pv-healthcare-financial-viability-colombia

Description

This dataset supports the article "Predictive Machine Learning Models for the Financial Viability of Sustainable Projects: A Case Study in Healthcare Facilities in Colombia." It provides energy consumption data, a financial cash-flow model, and dual Machine Learning results (binary classification of financial viability and continuous regression of Net Present Value) for 56 healthcare facilities in Barranquilla, Colombia. Consumption data originate from 2018 monthly billing records for 56 facilities (hospitals, CAMINO centers, and PASO primary-care centers). Electricity tariffs were updated to the values published by the regional utility Air-e for the 2025-2026 period (907, 943, 796, and 796 COP/kWh across the verified tariff checkpoints). For each facility, a 25-year discounted cash-flow model estimates CAPEX, annual O&M cost, avoided energy cost, Net Present Value (NPV), and Internal Rate of Return (IRR) under self-consumption solar PV. The repository includes: (1) the processed facility-level dataset with all technical-financial variables and the two ML target variables (Viabilidad_Financiera, VAN_COP); (2) PCA-reduced training and test sets used for the classifier and regressor; (3) classifier outputs (classification report, PCA feature importances); (4) regressor outputs (R²/MAE/RMSE metrics, PCA feature importances); (5) six figures covering correlation analysis, PCA explained variance, and exploratory consumption visualizations; and (6) the full manuscript and the MDPI citation style file. Note on traceability: the included classifier/regressor metrics correspond to a small validation split (45 train / 12 test) delivered with this data release. They differ from the final metrics reported in the article body (AUC = 0.993, F1 "not viable" class = 0.744, regression R² = 0.743), which were obtained using 5-fold GroupKFold cross-validation grouped by facility over the full 5,400-scenario Monte Carlo simulation. This dataset is therefore best understood as the facility-level processed data and a representative ML run, not a byte-for-byte reproduction of the article's headline metrics, unless a future version adds the complete Monte Carlo scenario file. Keywords: machine learning; solar photovoltaic; financial viability; Monte Carlo simulation; net present value; Gradient Boosting; XGBoost; healthcare facilities; energy poverty; Colombia; Barranquilla; Air-e tariff. License: CC BY 4.0. This dataset does not constitute financial, technical, or investment advice; it is intended for academic and research purposes only.

Steps to reproduce

1. Source data preparation: Monthly 2018 electricity consumption records were collected for 56 healthcare facilities in Barranquilla, Colombia, classified into three facility types: HOSPITAL, CAMINO, and PASO. Consumption (kWh/year) was held fixed as the baseline; tariffs (COP/kWh) were updated using Air-e's verified 2025-2026 published rates. 2. Financial model construction: For each facility, a 25-year discounted cash-flow model was built using NPV = -CAPEX + Sum[(Savings0 x (1-d)^(t-1) - OPEX) / (1+r)^t], for t = 1...25, where Savings0 = annual consumption x tariff x self-consumption ratio. Reference assumptions: solar irradiance 5.5 peak-sun-hours/day, performance ratio 80%, CAPEX $2.65M-$4.2M COP/kWp, annual O&M 1.5% of CAPEX, 25-year system life, 0.5% annual panel degradation, and a 12% base discount rate. 3. Monte Carlo simulation: 100 scenarios per facility (5,400 total) were generated by sampling five uncertain parameters: discount rate (triangular 8%/12%/22%), CAPEX multiplier (truncated normal, mean 1.0, sd 0.15, range 0.70-1.40), annual degradation (uniform 0.3%-0.8%), self-consumption ratio (truncated normal by facility type, sd 0.08, range 0.40-0.98), and tariff scenario multiplier (triangular 0.85/1.00/1.25). 4. Dimensionality reduction: The technical-financial input variables were standardized and reduced via Principal Component Analysis (PCA). The resulting components (PC1, PC2, PC3) are provided in classifier/train_data_pca.csv and classifier/test_data_pca.csv, retaining the original binary viability label and the continuous NPV target. 5. Model training: Two Gradient Boosting models (XGBoost where available) were trained: a binary classifier for financial viability and a regressor for NPV. Validation used GroupKFold (5 folds) grouped by facility to prevent data leakage between training and test sets. Class imbalance (majority "viable" class) was addressed via decision-threshold optimization on the precision-recall curve of the minority "not viable" class, rather than class weighting. 6. Outputs included in this release: classifier/classification_report.csv and regressor/regression_metrics.csv report performance on a representative train/test split (45/12 facilities) included with this data package; feature_importances_pca.csv files report PCA-component importance for each model; figures/ contains the correlation matrix, PCA explained-variance plot, and exploratory consumption charts. 7. To reproduce the article's headline metrics (AUC = 0.993, F1 = 0.744, R² = 0.743), replicate steps 1-5 over the full 5,400-scenario Monte Carlo dataset using 5-fold GroupKFold cross-validation grouped by facility, then apply threshold optimization as described in step 5. The full Monte Carlo scenario file is not included in this version; see the associated article for complete methodological detail.

Institutions

• Corporación Universitaria Iberoamericana

  Bogota D.C., Bogotá

Categories

Licence